What the documents look like
Parcel handoff happens at the point where a package leaves the warehouse, and the carrier scan carries responsibility. In store service and internal training work, staff need to understand where four document types come from: electronic waybills, handoff checklists, pickup scan records, and exception item registers. These materials update per batch and per day, so any citation must state the batch and version scope.
A training example that treats one batch's scan record as a general rule will lead trainees to misjudge responsibility. This page addresses citation source and traceability only. It does not cover waybill printing operations or pickup scan execution.
The knowledge base search node can select one or more knowledge bases that share the same vector model, and it outputs a citation array whose length may be zero. An empty array means the search path completed with no cited content. It does not mean the workflow terminated with an error. Citations must be checked item by item against source text. Answer fluency or a thinking process cannot substitute for source evidence. Semantic recall and relevance scores reflect similarity only; they do not filter permissions or isolate entities. Retrieval is not authorization, and values collected by a form do not mean business approval is complete. Hybrid retrieval merges multiple result paths through RRF; this is a fusion process, not a toggle to switch on. Image retrieval requires image input and model capability, and reranking requires a model enabled through its settings. Both are optional paths.
Hypothetical drill sample: a store training lead drafts a handoff statement: "According to item 7 of the March 12, 2025 handoff checklist, all packages in this batch were scanned and picked up by the carrier, with zero exception items." The source text provided for comparison reads: "March 12, 2025 handoff checklist, item 7: 120 items in this batch, 118 scanned and picked up by the carrier, 2 exception items registered, pending carrier confirmation." The question is whether the drafted sentence matches the source. The source names the subject (carrier), the object (120 packages in this batch), the status (118 scanned, 2 exceptions pending confirmation), and the effective scope (that batch on March 12, 2025).
Constraints at the citation tracing step
The search node's citation output is an array, and a zero-length array is a valid output that still passes the link onward. When the array is empty, the downstream step receives no citation content, and this gap should be recorded rather than classified as a failure. Each citation must be compared against its source text for the same subject, the same period, and the same field. A contradiction exists only when the same field holds explicitly mutually exclusive values. A value not stated in the source, or supported only by a related stage, is unsupported or pending verification.
Relevance scores and semantic recall do not act as permission filters or entity isolation. A low-scoring record may be filtered while an unauthorized record can still be recalled, so permission must be checked separately. Knowledge bases can be associated together only when they share the same vector model. If model identifiers differ, the mismatched bases are excluded and the actual observation is recorded; cross-model retrieval is not run for comparison.
The overall retrieval flow lists optional paths that do not all execute on every request. Image retrieval needs a corresponding image input and model capability. Reranking needs a model and must be enabled through its settings. Thinking budget is a reasoning budget and is not a citation token limit. RRF is a merge process across result paths, not a switch.
Configuration and implementation checks
| Configuration item | Setting or check for this page | Hypothetical trial starting point and reason |
|---|---|---|
| Associated knowledge bases | Check selected bases and material versions; only bases with the same vector model can be associated | Two bases with the same vector model are assumed, because cross-model association is not permitted |
| Search mode | Semantic retrieval, full-text retrieval, or hybrid retrieval | Hybrid retrieval is assumed, because the sample contains an item number and natural language |
| Citation limit | Maximum tokens cited per search | A starting point of 2000 tokens is assumed, to be compared against the sample source length |
| Minimum relevance | A 0-1 value effective only under semantic retrieval or reranking | A starting point of 0.3 is assumed, to be compared against the sample recall score |
| Question optimization | Coreference resolution and question expansion, adding one model call | Off is assumed, because the sample question already names the subject and date |
| Result reranking | Second-pass ordering with a ReRank model | On is assumed; model availability and the setting must be confirmed |
Three common errors
- Treating retrieval output as authorized execution. The input is the citation array returned by the search node, the comparison object is the business approval status, and the observable result is that citation content exists while the approval record is absent. Handoff confirmation must not proceed in that state.
- Treating an empty citation array as a system error. The input is the search node output
[], the comparison object is the workflow termination condition, and the observable result is that the link continues while the downstream step receives no citation. The gap should be recorded, not judged a failure. - Treating minimum relevance as a permission filter. The input is the relevance threshold setting, the comparison object is record access permission, and the observable result is that low-scoring records are filtered while unauthorized records can still be recalled. Permission must be checked separately.
How to verify the setup
- Compare the hypothetical sample source text word by word against the drafted sentence. The input is the handoff checklist source text and the drafted sentence above, the comparison object is "118 scanned, 2 exceptions" against "all scanned, zero exceptions," and the observable result is that the two do not match. The drafted sentence is unsupported.
- Use a paper input of
[]to check the downstream handling design. The input is an explicit empty citation array, the comparison object is whether the downstream node continues execution, and the observable result is that the link completes with no citation content. This confirms the empty result is not treated as an error. - Check the vector model identifiers of the knowledge bases. The input is the list of selected knowledge bases, the comparison object is each base's vector model name, and the observable result is that only bases with the same vector model are retained. Mismatched bases are excluded and the actual observation is recorded; cross-model retrieval is not run for comparison.